Thinking Machines Lab releases Inkling, 975-billion parameter open-weights AI model

Thinking Machines Lab, founded in February 2025 by former OpenAI executives Mira Murati, John Schulman, and Lilian Weng, has released Inkling, a 975-billion-parameter open-weight AI model. Supported by a record-breaking $2 billion seed round valuing the startup at $12 billion, the Mixture-of-Experts model features 41 billion active parameters and native multimodal capabilities. Inkling introduces controllable thinking effort to optimize token efficiency and ranks ninth on the Design Arena Agentic Web Dev leaderboard.

Inkling model release

• ▪Thinking Machines Lab was founded in February 2025 by former OpenAI executives Mira Murati, John Schulman, and Lilian Weng.
• ▪Thinking Machines Lab secured a $2 billion seed funding round at a $12 billion valuation.
• ▪Thinking Machines Lab released Inkling, a 975-billion-parameter open-weight artificial intelligence model, on July 15, 2026.

Mixture-of-Experts architecture

• ▪Inkling is a Mixture-of-Experts transformer model with 975 billion total parameters and 41 billion active parameters.
• ▪Inkling's Mixture-of-Experts design features 256 routed experts and 2 shared experts, with 6 routed experts active per token.

Controllable thinking effort

• ▪Inkling achieves the same performance as Nemotron 3 Ultra on the Terminal Bench 2.1 benchmark using approximately one-third of the tokens.
• ▪Inkling supports controllable thinking effort, allowing developers to adjust the model's thinking parameters to balance performance with token efficiency.

Multimodal capabilities

• ▪Inkling processes audio inputs as discrete dMel spectrograms and images as 40x40 pixel patches using an encoder-free architecture.
• ▪Inkling natively supports text, image, and audio input modalities, and it was pretrained on 45 trillion tokens of multimodal data.

Agentic coding capabilities

• ▪Inkling ranks ninth on the Design Arena Agentic Web Dev leaderboard with an Elo rating of 1257, placing it among the strongest open-weight models.
• ▪Inkling scores an Elo of 1238 on GDPval-AA v2 and 24% on the Tau 3 Banking benchmark, outperforming Kimi K2.6 and DeepSeek v4 Flash max.

Model calibration

• ▪Inkling was trained for calibration using reinforcement learning against proper scoring rules on a large corpus of resolved real-world questions.
• ▪Inkling's post-training utilized a rubric grader and a claims grader that performs agentic web searches to verify factual claims and reduce hallucinations.

3 sources